U.S. patent number 10,657,330 [Application Number 15/744,301] was granted by the patent office on 2020-05-19 for information extraction method and apparatus.
This patent grant is currently assigned to BOE TECHNOLOGY GROUP CO., LTD.. The grantee listed for this patent is BOE TECHNOLOGY GROUP CO., LTD.. Invention is credited to Zhenzhong Zhang.
View All Diagrams
United States Patent |
10,657,330 |
Zhang |
May 19, 2020 |
Information extraction method and apparatus
Abstract
The present invention is related to an information extraction
method. The information extraction method may comprise providing r
semantic relationships, acquiring entity pairs corresponding to the
semantic relationships, acquiring first instances based on the
entity pairs, and eliminating instances that do not have the
semantic relationships from the first instances to obtain second
instances. r is a positive integer. Each of the entity pairs
contains a pair of named entities. The first instances are
sentences containing the entity pairs.
Inventors: |
Zhang; Zhenzhong (Beijing,
CN) |
Applicant: |
Name |
City |
State |
Country |
Type |
BOE TECHNOLOGY GROUP CO., LTD. |
Beijing |
N/A |
CN |
|
|
Assignee: |
BOE TECHNOLOGY GROUP CO., LTD.
(Beijing, CN)
|
Family
ID: |
62024297 |
Appl.
No.: |
15/744,301 |
Filed: |
July 6, 2017 |
PCT
Filed: |
July 06, 2017 |
PCT No.: |
PCT/CN2017/091999 |
371(c)(1),(2),(4) Date: |
January 12, 2018 |
PCT
Pub. No.: |
WO2018/076774 |
PCT
Pub. Date: |
May 03, 2018 |
Prior Publication Data
|
|
|
|
Document
Identifier |
Publication Date |
|
US 20190005026 A1 |
Jan 3, 2019 |
|
Foreign Application Priority Data
|
|
|
|
|
Oct 28, 2016 [CN] |
|
|
2016 1 0972874 |
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06F
40/30 (20200101); G06F 40/295 (20200101) |
Current International
Class: |
G06F
40/30 (20200101); G06F 40/295 (20200101) |
References Cited
[Referenced By]
U.S. Patent Documents
Foreign Patent Documents
|
|
|
|
|
|
|
1319836 |
|
Oct 2001 |
|
CN |
|
102236692 |
|
Nov 2011 |
|
CN |
|
104281645 |
|
Jan 2015 |
|
CN |
|
105160046 |
|
Dec 2015 |
|
CN |
|
105550190 |
|
May 2016 |
|
CN |
|
105938495 |
|
Sep 2016 |
|
CN |
|
Other References
International Search Report dated Oct. 10, 2017, issued in
counterpart International Application No. PCT/CN2017/091999 (12
pages). cited by applicant .
Office Action dated Nov. 21, 2019, issued in counterpart CN
application No. 2016109728742, with English translation. (16
pages). cited by applicant.
|
Primary Examiner: Le; Thuykhanh
Attorney, Agent or Firm: Westerman, Hattori, Daniels &
Adrian, LLP
Claims
What is claimed is:
1. An information extraction method, comprising: providing r
semantic relationships, acquiring entity pairs corresponding to the
semantic relationships, acquiring first instances based on the
entity pairs, and eliminating instances that do not have the
semantic relationships from the first instances to obtain second
instances, wherein r is a positive integer, each of the entity
pairs contains a pair of named entities, and the first instances
are sentences containing the entity pairs; wherein eliminating the
instances that do not have the semantic relationships from the
first instances to obtain the second instances comprises:
extracting first features from each of the first instances based on
the entity pairs to construct a first instance-feature matrix
M.sub.nf, and constructing a semantic relationship-first instance
matrix M.sub.m, wherein the first features comprise lexical
features and/or syntactic features, n is a number of the first
instances, and f is a total number of all the first features; and
wherein eliminating the instances that do not have the semantic
relationships from the first instances to obtain the second
instances further comprises: obtaining a semantic
relationship-feature matrix M.sub.rf, wherein
M.sub.rf=M.sub.rn*M.sub.nf, decomposing the semantic
relationship-feature matrix M.sub.rf into M.sub.rk*M.sub.kf by a
nonnegative matrix factorization method, obtaining M.sub.nk by
multiplying M.sub.nf by M.sub.kf.sup.T, obtaining similarity
between each of the first instances and all the semantic
relationships respectively based on M.sub.nk and M.sub.rk.sup.T,
and screening out the second instances from the first instances
based on the similarity, wherein M.sub.rk is a representation
matrix of the semantic relationships in a latent semantic space and
M.sub.nk is a representation matrix of the first instances in the
latent semantic space.
2. The information extracting method according to claim 1, wherein
acquiring the entity pairs corresponding to the semantic
relationships and acquiring the first instances based on the entity
pairs comprises: acquiring the entity pairs corresponding to the
semantic relationships from a knowledge base; tagging sentences
that contain the named entities of all the entity pairs in a
database using a named entity recognition tool; and retrieving the
first instances containing the entity pairs from the tagged
sentences.
3. The information extracting method according to claim 1, wherein
obtaining the similarity between each of the first instances and
all the semantic relationships respectively based on M.sub.nk and
M.sub.rk.sup.T comprises: obtaining the similarity between each of
the first instances and all the semantic relationships respectively
through cosine similarity based on M.sub.nk and M.sub.rk.sup.T.
4. The information extracting method according to claim 1, wherein
screening out the second instances from the first instances based
on the similarity comprises: normalizing the similarity so that a
sum of the normalized similarity of each of the first instances
with all of the semantic relationships respectively is 1, obtaining
an information entropy of each of the first instances based on the
normalized similarity of each of the first instances with all of
the semantic relationships respectively, and selecting the first
instances whose information entropy is less than a predetermined
threshold as the second instances.
5. The information extracting method according to claim 1, further
comprising: extracting second features from the second instances
based on the entity pairs to train a classifier, wherein the second
features comprise lexical features and/or syntactic features.
6. The information extracting method according to claim 5, further
comprising: identifying semantic relationships from text sentences
and classifying the text sentences based on the semantic
relationships using the classifier.
7. The information extracting method according to claim 6, wherein
identifying the semantic relationships from the text sentences and
classifying the text sentences using the classifier comprises:
using the named entity recognition tool to tag sentences that
contain the named entities in a database, retrieving text sentences
containing the entity pairs from the tagged sentences, and
identifying semantic relationships of the entity pairs in the text
sentences according to the classifier, and classifying the text
sentences based on the semantic relationships according to the
classifier.
8. The information extracting method according to claim 1, wherein
the lexical features include at least one selected from the group
consisting of entity pair's position in a sentence, word sequence
between the entity pairs, characteristic sequence between the
entity pairs, left window of size X of the entity pair, and right
window of size X of the entity pair, wherein X.gtoreq.0, and the
syntax features include at least one selected from the group
consisting of the shortest dependency path between the entity
pairs, left window of size Y of the entity pair, and right window
of size Y of the entity pair, wherein Y.gtoreq.0.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
This application claims the benefit of the filing date of Chinese
Patent Application No. 201610972874.2 filed on Oct. 28, 2016, the
disclosure of which is hereby incorporated by reference.
TECHNICAL FIELD
This invention relates to a smart apparatus technology, and more
particularly, to an information extraction method and
apparatus.
BACKGROUND
In most technical fields, most of the knowledge comes from
unstructured text. With rapid development of various fields,
research literature expands exponentially and enriches relevant
knowledge in various fields.
However, the exploding growth of research literature goes far
beyond a speed of people analyzing the literature. It is very
difficult for researchers to find needed information quickly from a
literature library. Therefore, helping researchers to extract
useful information quickly from massive literature has attracted
wide attention from academia and industry.
BRIEF SUMMARY
Accordingly one example of the present disclosure is an information
extraction method. The information extraction method may comprise
providing r semantic relationships, acquiring entity pairs
corresponding to the semantic relationships, acquiring first
instances based on the entity pairs, and eliminating instances that
do not have the semantic relationships from the first instances to
obtain second instances. r is a positive integer. each of the
entity pairs contains a pair of named entities, and the first
instances are sentences containing the entity pairs.
Acquiring the entity pairs corresponding to the semantic
relationships and acquiring the first instances based on the entity
pairs may comprise acquiring the entity pairs corresponding to the
semantic relationships from a knowledge base; tagging sentences
that contain the named entities of all the entity pairs in a
database using a named entity recognition tool; and retrieving the
first instances containing the entity pairs from the tagged
sentences.
Eliminating the instances that do not have the semantic
relationships from the first instances to obtain the second
instances may comprise extracting first features from each of the
first instances based on the entity pairs to construct a first
instance-feature matrix M.sub.nf, and constructing a semantic
relationship-first instance matrix M.sub.m. The first features may
comprise lexical features and/or syntactic features. n is a number
of the first instances and f is a total number of all the first
features.
Eliminating the instances that do not have the semantic
relationships from the first instances to obtain the second
instances may further comprise obtaining a semantic
relationship-feature matrix M.sub.rf, wherein
M.sub.rf=M.sub.m*M.sub.nf, decomposing the semantic
relationship-feature matrix M.sub.rf into M.sub.rk*M.sub.kf by a
nonnegative matrix factorization method, obtaining M.sub.nk by
multiplying M.sub.nf by M.sub.kf.sup.T, obtaining similarity
between each of the first instances and all the semantic
relationships respectively based on M.sub.nk and M.sub.rk.sup.T,
and screening out the second instances from the first instances
based on the similarity. M.sub.rk is a representation matrix of the
semantic relationships in a latent semantic space and M.sub.nk is a
representation matrix of the first instances in the latent semantic
space.
Obtaining the similarity between each of the first instances and
all the semantic relationships respectively based on M.sub.nk and
M.sub.rk.sup.T may comprise obtaining the similarity between each
of the first instances and all the semantic relationships
respectively through cosine similarity based on M.sub.nk and
M.sub.rk.sup.T.
Screening out the second instances from the first instances based
on the similarity may comprise normalizing the similarity so that a
sum of the normalized similarity of each of the first instances
with all of the semantic relationships respectively is 1, obtaining
an information entropy of each of the first instances based on the
normalized similarity of each of the first instances with all of
the semantic relationships respectively, and selecting the first
instances whose information entropy is less than a predetermined
threshold as the second instances.
The information extracting method may further comprise extracting
second features from the second instances based on the entity pairs
to train a classifier, wherein the second features comprise lexical
features and/or syntactic features. The information extracting
method may further comprise identifying semantic relationships from
text sentences and classifying the text sentences based on the
semantic relationships using the classifier.
Identifying the semantic relationships from the text sentences and
classifying the text sentences using the classifier may comprise
using the named entity recognition tool to tag sentences that
contain the named entities in a database, retrieving text sentences
containing the entity pairs from the tagged sentences, and
identifying semantic relationships of the entity pairs in the text
sentences according to the classifier, and classifying the text
sentences based on the semantic relationships according to the
classifier.
The lexical features may include at least one selected from the
group consisting of entity pair's position in a sentence, word
sequence between the entity pairs, characteristic sequence between
the entity pairs, left window of size X of the entity pair, and
right window of size X of the entity pair, wherein X.gtoreq.0. The
syntax features may include at least one selected from the group
consisting of the shortest dependency path between the entity
pairs, left window of size Y of the entity pair, and right window
of size Y of the entity pair, wherein Y.gtoreq.0.
Another example of the present disclosure is an information
extracting apparatus. The information extracting apparatus may
comprise an acquisition module and a screening module. The
acquisition module may be configured to obtain entity pairs
corresponding to a given semantic relationship and acquire first
instances based on the entity pairs. The screening module may be
configured to remove instances that do not have the given semantic
relationship from the first instances to obtain second instances.
The first instances are sentences containing the entity pairs and
each of the entity pairs contains a pair of named entities.
That the acquisition module is configured to obtain entity pairs
corresponding to a given semantic relationship and acquire first
instances based on the entity pairs may comprise that the
acquisition module is configured to obtain the entity pairs
corresponding to the given semantic relationship from a knowledge
base, use a named entity recognition tool to tag sentences in a
database that contain the named entities of all the entity pairs,
and retrieve the first instances containing the entity pairs from
the tagged sentences.
That the screening module is configured to remove instances that do
not have the given semantic relationship from the first instances
to obtain second instances may comprise that the screening module
is configured to extract first features from each of the first
instances based on the entity pairs, construct a first
instance-feature matrix M.sub.nf, and construct a semantic
relationship-first instance matrix M.sub.rn, wherein n is a number
of the first instance, f is a number of all the first features, and
the first features comprise lexical features and/or syntactic
features.
That the screening module is configured to remove instances that do
not have the given semantic relationship from the first instance to
obtain second instances may further comprise that the screening
module is configured to obtain a semantic relationship-feature
matrix M.sub.rf, wherein M.sub.rf=M.sub.rn*M.sub.nf, decompose the
semantic relationship-feature matrix M.sub.rf into
M.sub.rk*M.sub.kf by a nonnegative matrix factorization method,
wherein k<f, obtain M.sub.nk by multiplying M.sub.nf by
M.sub.kf.sup.T, and obtain similarity between each of the first
instances and all the semantic relationships respectively based on
M.sub.nk and M.sub.rk.sup.T, and screen out the second instances
from the first instances based on the similarity. r is a number of
the semantic relationships, M.sub.rk is a representation matrix of
the semantic relationships in a latent semantic space, and M.sub.nk
is a representation matrix of the first instances in the latent
semantic space.
Obtain the similarity between each of the first instances and all
the semantic relationships respectively based on M.sub.nk and
M.sub.rk.sup.T may comprise obtain the similarity between each of
the first instances and all the semantic relationships respectively
through cosine similarity based on M.sub.nk and M.sub.rk.sup.T.
Screen out the second instances from the first instances based on
the similarity may comprise normalize the similarity so that a sum
of nonnalized similarity of each of the first instances and all of
the semantic relationships respectively is 1, obtain an information
entropy of each of the first instances based on the normalized
similarity of each of the first instances and all of the semantic
relationships respectively, and select the first instances whose
information entropy is less than a predetermined threshold as the
second instances.
The information extracting apparatus may further comprise a
classifier training module and an information extraction module.
The classifier training module may be configured to extract first
features from the second instances based on the entity pairs to
train the classifier. The information extraction module may be
configured to identify semantic relationships from text sentences
using the classifier and classify the text sentences based on the
semantic relationships using the classifier.
That the information extraction module is configured to identify
semantic relationships from text sentences and classify the text
sentences using the classifier may comprise that the information
extraction module is configured to use the named entity recognition
tool to tag sentences in a database that contain named entities of
the entity pairs, retrieve text sentences containing the entity
pairs from the tagged sentences, identify semantic relationships
based on the entity pairs in the instances using the classifier,
and classify the text sentences based on the semantic relationships
using the classifier.
The lexical features may include at least one selected from the
group consisting of entity pair's position in a sentence, word
sequence between the entity pairs, characteristic sequence between
the entity pairs, left window of size X of the entity pair, and
right window of size X of the entity pair, wherein X.gtoreq.0. The
syntax features may include at least one selected from the group
consisting of the shortest dependency path between the entity
pairs, left window of size Y of the entity pair, and right window
of size Y of the entity pair, wherein Y.gtoreq.0.
BRIEF DESCRIPTION OF THE DRAWINGS
The subject matter which is regarded as the invention is
particularly pointed out and distinctly claimed in the claims at
the conclusion of the specification. The foregoing and other
objects, features, and advantages of the invention are apparent
from the following detailed description taken in conjunction with
the accompanying drawings in which:
FIG. 1 is a flow chart of an information extraction method
according to an embodiment.
FIG. 2 is a flow chart of a method of obtaining first instances
containing entity pairs from a database according to an
embodiment.
FIG. 3 is a flow chart of a method of screening out second
instances from the first instances according to an embodiment.
FIG. 4 is a flow chart of a method of screening out second
instances from the first instances based on similarity according to
an embodiment.
FIG. 5 is a flow chart of an information extraction method
according to an embodiment.
FIG. 6 is a flow chart of a method of classifying semantic
relationships using a classifier according to an embodiment.
FIG. 7 is a schematic diagram of an information extracting
apparatus according to an embodiment.
FIG. 8 is a schematic diagram of an information extracting
apparatus according to an embodiment.
DETAILED DESCRIPTION
The present invention is described with reference to embodiments of
the invention. Throughout the description of the invention
reference is made to FIGS. 1-8. When referring to the figures, like
structures and elements shown throughout are indicated with like
reference numerals. It will be apparent that the described
embodiments are merely part of the embodiments and are not intended
to be exhaustive example. All other embodiments obtained by those
of ordinary skill in the art without making creative work are
within the scope of the present disclosure.
This invention relates to a relation extraction method, which aims
to identify and categorize relations between pairs of entities in
text. Traditional relation extraction methods usually need a lot of
annotated data as training corpus. However, annotating data by
human is time-consuming and expensive.
To address this limitation, the present invention presents a method
which can automatically gather labeled data by heuristically
aligning entities in text with those in a knowledge base. However,
this method often results in a noisy training corpus because
sentences that mention two entities are likely to express different
relations in a knowledge base. For example, if entities "Steve
Jobs" and "Apple" have "CEO of" relation in a knowledge base, then
the method will extract all sentences containing both entities,
e.g.
"Steve Jobs is the CEO of Apple" and "Steve Jobs founded Apple in
1976", from a given text corpus. But, in fact, "Steve Jobs is the
CEO of Apple" expresses the relation "CEO of" and "Steve Jobs
founded Apple in 1976" expresses the relation "Founder of". So
"Steve Jobs founded Apple in 1976" is a noisy training instance for
relation "CEO of". To address this problem, the present invention
improves the above method by identifying reliable instances from
noisy instances by inspecting whether an instance is located in a
semantically consistent region. Specifically, given some training
instances, the method first models each relation type as a linear
subspace spanned by its training instances. Then, the local
subspace around an instance is modeled and characterized by seeking
the sparsest linear combination of training instances which can
reconstruct the instance.
In summary, compared with traditional methods of manual labeling,
the method present by the present invention can automatically
gather labeled data by considering the semantic consistence between
gathered instances and given training instances, which can save a
lot of time and labor costs. Thus, the present invention improves
efficiency and accuracy, and reduces cost relative to the
traditional methods.
Specifically, the present invention has at least the following
advantages:
First, the present invention uses supervised learning methods,
which improves accuracy of the extraction.
Second, compared with the traditional semantic relation extraction
method based on supervised learning, the present invention uses an
automatic annotation method of training corpus based on matrix
decomposition, which can automatically label a large amount of
training data based on a given knowledge base, thereby reducing
annotation cost. In the field of machine learning, supervised
learning requires manual training of certain training data, but the
cost of manual labeling is very high. So the present invention has
great advantages in obtaining a fairly accurate model with less
manual annotation. The basic concept of the present invention is to
use the relevant knowledge in an existing knowledge base to
automatically label the training data, thereby reducing the cost of
manual labeling
Furthermore, the matrix representation is used to compute the
semantic consistence between gathered instances and some given
training instances.
FIG. 1 shows a flow chart of an information extraction method
according to an embodiment. As shown in FIG. 1, the information
extraction method comprises step S10 and step S20.
In step S10, semantic relationships are first provided. Then,
entity pairs corresponding to the semantic relationships are
obtained. Then, first instances are obtained based on the entity
pairs. The first instances comprise sentences containing the entity
pairs.
The semantic relationships and the entity pairs are not limited to
any particular technical field as long as there is a database of
the semantic relationships and a database of the entity pairs
corresponding to the semantic relationships. Here the database can
be selected from existing databases or newly made databases based
on needs. Embodiments of the present disclosure use examples in a
medical field merely for purpose of illustration. In addition,
specific methods for obtaining the first instances are not limited
as long as the first instances obtained comprise sentences
containing the entity pairs.
In step S20, instances that do not have the given semantic
relationships are removed from the first instances to form second
instances. That is, sentences in the first instance that do not
have the given semantic relationships are removed. Normally
sentences that do not have the given semantic relationships are
called noise data. After the noise data is removed from the first
instances, second instances are obtained. Specific method of
removing the noise data is not limited herein.
An example of the present disclosure is an information extraction
method. According to given semantic relationships, entity pairs
corresponding to the semantic relationships are obtained. First
instances are obtained based on the entity pairs. The first
instances are screened with the given semantic relationships to get
second instances. That is, sentences that do not have the given
semantic relationships are removed from the first instances. As
such, noise data is eliminated from the sample data. In this way,
when the sample data is used to train a model, accuracy of the
model can be improved, and accordingly accuracy of acquired
information can be improved.
In one embodiment, as shown in FIG. 2, step S10 includes step S11
to step S13.
In step S11, according to given semantic relationships, entity
pairs corresponding to the semantic relationships are obtained from
a knowledge base.
In the medical field, the Unified Medical Language System (UMLS)
consists of four parts: metathesaurus, semantic network,
information sources map, and specialist lexicon. In one embodiment,
the given semantic relationships comprise at least one of the 54
semantic relationships classified in the semantic network.
The knowledge base may be a pre-selected existing knowledge base.
For example, it may be a knowledge base normally used by an
ordinary skill person in the art such as UMLS.
An entity pair refers to two entities that have a certain semantic
relationship in the knowledge base. The entity may be, for example,
a disease, a drug, a gene, or the like. The entity pair referred to
in the embodiment of the present disclosure may be a plurality of
entity pairs, as long as each entity pair corresponds to a given
semantic relationship.
The database is a database containing text, for example, an
existing database such as CNKI (China National Knowledge Internet)
or Wanfang. Here, the type of the text is not limited. For example,
it can be papers, journals, books, and so on.
In addition, in embodiments of the present disclosure, there is one
or more first instances containing the entity pairs obtained from
the database, and not each of the instances expresses the given
semantic relationship. For example, a semantic relationship given
is "treatment". It is found from the knowledge base that entity
pairs having this semantic relationship are "metformin" and
"gestational diabetes", "aspirin" and "cold". Instance 1 retrieved
is "aspirin can be used to treat the cold." Instance 2 retrieved is
"scientists claim that aspirin is one of the culprits of the 1981
cold epidemic." Instance 3 retrieved is "metformin is the first
choice for the treatment of gestational diabetes." The semantic
relationships between the entity pairs in Instance 1 and Instance 3
can be read as "treatment". However, In Instance 2, the semantic
relationship between the entity pair can be read as "caused" rather
than "treated." Thus, Instance 2 is noise data and removed. Only
Instance 1 and Instance 3 are screened out as second instances.
In step S12, using a Named Entity Recognition tool, the sentences
containing the named entities in a database are tagged.
The Named Entity Recognition (NER) refers to entities that have
specific meanings in the recognition text, which mainly include
names of persons, names of places, organization names, proper
nouns, drug names, gene names, and disease names, and so on. When
using a named entity recognition tool, entities with the
above-mentioned specific meanings contained in the sentence are
tagged. By tagging entities in sentences, the sentences containing
the named entities are tagged.
In step S13, the first instances containing the entity pairs are
retrieved from the tagged sentences.
That is, the sentences containing the given entity pairs (i.e., the
entity pairs acquired in S11) are retrieved from only the sentences
tagged in step S12. Other sentences are removed. If the sentence
contains only one entity in a given entity pair, the sentence is
also removed.
The embodiment of the present disclosure obtains first instances
containing entity pairs from a database by using a named entity
recognition tool. The named entity recognition tool is a mature
technology. Furthermore, this method is efficient and has low
cost.
In one embodiment, as shown in FIG. 3, step S20 includes step S21
to step S27.
In step S21, first features are extracted from each of the first
instances based on the entity pairs to construct a first
instance-feature matrix M.sub.nf. n is the number of the first
instances, and f is the number of all the first features. The first
features include lexical features and/or syntactic features.
In one embodiment, first features are extracted from each of the
first instances according to the entity pairs. Extracting the first
features may comprise extracting lexical features and/or syntactic
features in each sentence of the first instances having the entity
pairs, and then taking intersection or union of the extracted
lexical features and/or syntactic features. Of course, there can be
other ways to extract.
The first instance-feature matrix M.sub.nf is constructed, that is,
whether each of the first instances has the extracted first
features is determined. In the first instance-feature matrix
M.sub.nf, each row of data may indicate whether the first instance
corresponding to the row has the extracted first features. For
example, "1" may indicate that the first instance has a first
feature, and "0" may indicate that the first instance does not have
a first feature. In the following embodiment, "1" represents the
first instance having a certain first feature, and "0" represent
the first instance without a certain first feature.
For example, there are four first instances, that is n=4. Six first
features are extracted, that is f=6. Instance 1 has a first and a
third first feature. Instance 2 has a first and a fourth first
feature. Instance 3 has a fifth and a sixth first feature. Instance
4 has a second and a third first feature. The obtained first
instance-feature matrix M.sub.nf is:
##EQU00001##
In step S22, a semantic relationship-first instance matrix M.sub.rn
is constructed. r is the number of semantic relationships.
In one embodiment, in the semantic relationship-first instance
matrix M.sub.rn, first, a plurality of semantic relationships is
given. Each column of data in the matrix may indicate whether the
first instance corresponding to the column belongs to any of the
semantic relations. For example, "1" means that the first instance
belongs to a semantic relationship, and "0" means that the first
instance does not belong to a semantic relationship.
For example, five semantic relationships are given, that is, r=5.
There are four first instances. Instance 1 and Instance 2 belong to
the first semantic relationship. Instance 1 and Instance 3 belong
to the second semantic relationship. Instance 1 and Instance 4
belong to the third semantic relationship. Instance 2 and Instance
3 belong to the fourth semantic relationship. Instance 3 and
Instance 4 belong to the fifth semantic relationship. Then the
obtained semantic relationship-first instance matrix M.sub.rn
is:
##EQU00002##
In step S23, a semantic relationship-feature matrix M.sub.rf is
constructed, wherein M.sub.rf=M.sub.rn*M.sub.nf. That is, the
semantic relationship-feature matrix M.sub.rf is obtained by
multiplying the semantic relationship-first instance matrix
M.sub.rn by the first instance-feature matrix M.sub.nf.
For example, based on the above examples, M.sub.54 and M.sub.46 are
multiplied to obtain a semantic relationship-feature matrix
M.sub.rf, which is
##EQU00003##
In step S24, the semantic relationship-feature matrix M.sub.rf is
decomposed into M.sub.rk*M.sub.kf by non-negative matrix
factorization (NMF), wherein k<f and M.sub.rk is a
representation matrix of the semantic relationships in a latent
semantic space.
The nonnegative matrix factorization (NMF) is to find two low-rank
non-negative matrixes M.sub.rk and M.sub.kf, so that
M.sub.rf=M.sub.rk*M.sub.kf. Non-negative matrix factorization
process can be realized by using MATLAB, C, C++ or other language
programs.
Here, the non-negative matrix factorization of the semantic
relationship-feature matrix M.sub.rf may be performed by filtering
the f first features extracted in step S21 to select k first
features which have relatively high repetition rates or are
relatively more important. That is, the semantic
relationship-feature matrix M.sub.rf is mapped into a latent
semantic space to obtain a representation matrix M.sub.rk of the
semantic relationships in the latent semantic space and a
representation matrix M.sub.kf.sup.T of the first features in the
latent semantic space. M.sub.rk and M.sub.kf are obtained by
non-negative matrix factorization of M.sub.rf. Then, M.sub.kf is
transposed to obtain M.sub.kf.sup.T. In one embodiment, in order to
improve desirability of samples, n, r, f are relatively large
numbers. As such, in practice, k may be much smaller than f. For
example, when k=3, M.sub.56 is decomposed into
M.sub.56=M.sub.53*M.sub.36.
In step S25, M.sub.nk is obtained by multiplying M.sub.nf by
M.sub.kf.sup.T. M.sub.nk is a representation matrix of the first
instances in the latent semantic space. That is, the first
instance-feature matrix M.sub.nf is multiplied by the
representation matrix of the first feature in the latent semantic
space M.sub.kf.sup.T to obtain the representation matrix M.sub.nk
of the first instance in the latent semantic space. For example,
M.sub.43=M.sub.46*M.sub.36.sup.T.
In step S26, similarity between each of the first instances and the
semantic relationships respectively are obtained based on M.sub.nk
and M.sub.rk.sup.T.
In one embodiment, M.sub.nk can be multiplied by M.sub.rk.sup.T to
obtain the representation matrix of the first instance-semantic
relationship in the latent semantic space. Each row of data in the
matrix represents the similarity between the first instance
represented by the row and each of the semantic relationships
respectively. Of course, other computing methods can also be used
to get the similarity between first instance and semantic
relationships.
For example, M.sub.nk is multiplied by M.sub.rk.sup.T to obtain a
representation matrix of the first instance-semantic relationship
in the latent semantic space, which is
##EQU00004## The first row of data in the matrix (2,3,7,4)
indicates that the similarity between Instance 1 and the first
semantic relationship is 2. The similarity between Instance 1 and
the second semantic relationship is 3. The similarity between
Instance 1 and the third semantic relationship is 7. The similarity
between Instance 1 and the fourth semantic relationship is 4.
Likewise, the similarity between each of the instances and each of
the semantic relationships respectively can be obtained.
In step S27, the first instances are screened based on the
similarity to obtain second instances. That is, based on the
similarity between each of the first instances and each of the
semantic relationships respectively calculated in step S26, the
first instances having obvious differences in terms of the
similarity are selected as the second instances. For example, in
the above-mentioned matrix
##EQU00005## representing the similarity between three first
instances and four semantic relationships, the similarity between
Instance 1 and the third semantic relationship is relatively high.
As such, it can be clearly determined that instance 1 belongs to
the third semantic relationship. The difference in terms of the
similarity between Instance 2 and the four semantic relationships
is not large. As such, it cannot be clearly determined that which
semantic relationship instance 2 belongs. Instance 3 has a high
similarity with the fourth semantic relationship. As such, it is
clearly determined that it belongs to the fourth semantic
relationship. Thus, Instances 1 and 3 of the first instance are
screened out as second instances.
In the embodiment of the present disclosure, the semantic
relationships, the instances, and the features are first mapped
into the latent semantic space by a nonnegative matrix
factorization method. Then, the similarity between each of the
first instances and the semantic relationships respectively is
obtained in the latent semantic space. Then, based on whether there
is clear difference in similarity, that is, whether it can be
clearly determined that an instance belongs to a certain semantic
relationship, the second instances are screened out from the first
instances. This data selection method based on matrix decomposition
has advantages such as small amount of calculation and high
efficiency during process of removing noise data.
In one embodiment, obtaining the similarity between each of the
first instances and the semantic relationships respectively is
obtained based on M.sub.nk and M.sub.rk.sup.T comprises:
The similarity between each of the first instances and the semantic
relationships respectively is obtained through cosine similarity
based on M.sub.nk and M.sub.rk.sup.T. The cosine similarity is to
use a cosine value of an angle of two vectors in a vector space as
a measure of difference between the two vectors. The closer the
cosine value is to 1, the closer the angle is to 0 degree, and the
more similar the two vectors are. In the present disclosure, each
row of data in M.sub.nk and each column of data in M.sub.rk.sup.T
are used as a vector respectively. The similarity between the first
instances and the semantic relationships is obtained by calculating
the cosine similarity.
In this embodiment, the similarity between each of the first
instances and the semantic relationships respectively is obtained
by calculation of the cosine similarity. This method not only
obtains excellent results, but also is quick, convenient, and
simple.
In one embodiment, as shown in FIG. 4, screening out second
instances from the first instances based on the similarity
comprises step S271 to step S273.
In step S271, the similarity is normalized so that a sum of
normalized similarity between each of the first instances and all
semantic relationships respectively is 1.
The normalization comprises: mapping a group of data in a
parenthesis to a group of numbers within a range of 0 and 1, and a
sum of the group of numbers becomes 1. Then, a ratio of each number
to the sum of the group of numbers is written down. By normalizing
the similarity, the similarity between each of the first instances
and the semantic relationships respectively can be considered as a
probability distribution.
For example, the similarity between Instance 1 and all semantic
relationships respectively calculated by step S26 is (2, 3, 7, 4).
The similarity between Instance 1 and all semantic relationships
respectively after normalization is
##EQU00006## That is (0.125, 0.1875, 0.4375, 0.25).
In step S272, information entropy of each of the first instances is
calculated based on the similarity of the first instance after the
normalization process. That is, the similarity between each of the
first instances and all semantic relationships respectively is
considered as a probability distribution. Information entropy of
each of the first instances is calculated according to calculation
formula of the information entropy. Information is an abstract
concept. Entropy represents a physical quantity of a state of a
material system, which indicates the extent to which the state may
occur. The information entropy
.function..times..times..function..times..times..times..function..functio-
n. ##EQU00007## is probability that the first instance belongs to a
semantic relationship, that is, the similarity between the
normalized first instance and each of the semantic relationships
respectively.
For example, the similarity between Instance 1 and all semantic
relationships respectively is (0.125, 0.1875, 0.4375, 0.25). Then
the information entropy of Instance 1 is H=-(0.125 log 0.125+0.1875
log 0.1875+0.4375 log 0.4375+0.25 log 0.25)=0.5567856.
In step S273, the first instances whose information entropy is
smaller than a predetermined threshold are selected as the second
instances.
The lower the information entropy, the more definite the
information contained in the first instance, as the instances which
have low information entropy are more favorable to the training
model. Thus, the first instances whose information entropy is
greater than the predetermined threshold are noise data.
Accordingly, the second instances are selected from the first
instances based on the information entropy of each of the first
instances, thereby removing the noise data.
In addition, a value of the predetermined threshold of the
information entropy is not limited herein, and the value can be
reasonably determined based on data quantities.
For example, if the predetermined threshold of the information
entropy is 0.6, the above Instance 1 having information entropy of
0.5567856 will be selected as the second instance. If the
predetermined threshold of the information entropy is 0.5, the
above Instance 1 having information entropy of 0.5567856 will not
be selected as the second instance. Instead, it will be removed as
the noise data.
In the embodiment of the present disclosure, the similarity is
normalized so that a sum of the similarity between each of the
first instances and all semantic relationships respectively is 1.
Then, the similarity is regarded as a probability distribution.
Information entropy of each of the instances is calculated. Then,
instances having information entropy larger than a predetermined
threshold are removed as noise data. As such, the second instances
are selected from the first instances. In the present embodiment,
amount of calculation thereof is small and acquisition is
convenient.
In one embodiment, as shown in FIG. 5, the information extraction
method further comprises step S30 and step S40.
In step S30, second features are extracted from each of the second
instances based on the entity pairs to train a classifier. A target
of the classifier is to classify based on given semantic
relationships. The second features comprise lexical features and/or
syntactic features.
Extracting the second features may comprise the following:
extracting lexical features and/or syntactic features of the entity
pairs in each sentence of the second instances, and then taking
intersection or union of the extracted lexical features and/or
syntactic features. Of course, there can be other ways to extract.
The classifier is trained using the extracted second features.
Here, types of lexical features and syntactic features are not
limited herein.
Furthermore, a conventional method of training the classifier can
be employed herein. The above extracted second features are used by
the classifier. A final outputted result of the classifier is to
classify text sentences into specific semantic relationships
respectively.
In step S40, according to the classifier, semantic relationships
are identified and text sentences are classified based on the
semantic relationships. That is, the trained classifier is applied
to a database to identify and classify semantic relationships
between the entity pairs in the text sentences.
In the embodiment of the present disclosure, a classifier is
trained by using the second instances (sample data) after
eliminating the noise data. As such, parameters of the classifier
trained by the second instances are more desirable, thereby
improving accuracy of extracted information from the text sentences
by the classifier.
In addition, through use of the trained classifier to supervise and
learn from existing databases from long-distance, the semantic
relationships between the entity pairs in the sentences are
identified and the sentences are classified automatically, instead
of manually tagging data. As such, cost of the tagging is reduced,
coverage of semantic relationship extraction is improved, and the
problem of sparse data is solved.
In one embodiment, as shown in FIG. 6, the above-described step S40
includes step S41 to step S43.
In step S41, a named entity recognition tool is used to tag
sentences containing the named entities in a database.
In step S42, text sentences containing the entity pairs are
retrieved from the tagged sentences.
In step S43, according to the classifier, semantic relationships of
the entity pairs in the text sentences are identified and
classified. That is, information extraction is performed for each
of the text sentences extracted in step S42, and a final result of
the output is which semantic relationship the entity pair in the
text sentence belongs. For example, "1" represents belonging to a
certain semantic relationship, and "0" represents not belonging to
a certain semantic relationship.
In the embodiment, sentences containing named entities in a
database are tagged by a named entity recognition tool. Then, text
sentences that contain the entity pairs are retrieved. Then, the
classifier is used to identify the corresponding semantic
relationships of the entity pairs in the text sentences and
classify the text sentences based on the semantic relationships.
This embodiment has advantages such as using mature technology, low
cost, and high efficiency.
In one embodiment, the lexical features include at least one of
entity pair's position in a sentence, word sequence between the
entity pair, characteristic sequence between the entity pair, left
window of size X of the entity pair X, or right window a size X of
the entity pair, wherein X.gtoreq.0.
The syntax features include at least one of the shortest dependency
path between the entity pairs, left window of size Y of the entity
pair, or right window of size Y of the entity pair, wherein
Y.gtoreq.0.
The entity pair's position in a sentence refers to whether the pair
of entities in the sentence is adjacent to each other or there are
other words between the entity pair. The word sequence between the
entity pair refers to the order of each of the entities located in
the sentence, respectively. The characteristic sequence between the
entity pair refers to characteristic of the words (such as nouns,
verbs, etc.) before and between the pair of entities. The left
window of size X of the entity pair refers to the number of words
from the entity in the front to the beginning of the sentence. The
right window of size X of the entity pair refers to the number of
words from the entity in the front to the entity in the back. The
shortest dependency path between the entity pair refers to the
number of words between the pair of entities. The left window of
size Y of the entity pair refers to number of words from the entity
in the front to the beginning of the sentence in a tree view listed
according to the syntax features. The right window of size Y of the
entity pair refers to number of words from the entity in the front
to the entity in the back in a tree view listed according to the
syntax features.
Here, a feature refers to one feature or a combination of multiple
features. When the feature refers to one feature, the text sentence
containing this feature means that the text sentence has the
feature. When the feature refers to a combination of two features,
only that the text sentence includes the two features means that
the text sentence is considered to contain the feature.
In addition, the X as the size of the left window of the entity
pair and the X as the size of the right window of the entity pair
can be the same or different. The Y as the size of the left window
of the entity pair and the Y as the size of the right window Y of
the entity pair can be the same or different. The values of X and Y
may be the same or different.
A feature can be used as a number of features by changing
parameters of the feature. For example, for the feature of left
window of size X of the entity pair, the left window of size 3 of
the entity pair, the left window of size 5 of the entity pair, the
left window of size 6 of the entity pair each can be used as a
feature.
For example, a retrieved text sentence is "scientists claim aspirin
is one of the culprits of the 1981 cold epidemic." The position of
the entity pair in the sentence is not adjacent to each other. If
the extracted feature is the entity pair's position in the sentence
being adjacent, the feature of the text sentence is "0." The word
sequence of the entity pair in the text sentence is that the first
entity is the third word and the second entity is the twelfth word.
If the extracted feature is exactly the same as the above feature,
the feature of the text sentence is "1." In the text sentence, the
characteristic sequence of the word between the entity pair is
"verb, numeral, preposition, determiner, noun, preposition,
determiner, numeral." If the extracted feature is different from
the above feature, the feature in the text sentence is "0." The
size of the left window of the entity pair in the text sentence is
three. If the extracted feature is the same as the above feature,
the feature in the text sentence is "1." The size of the right
window of the entity pair in the text sentence is zero, and if the
extracted feature is different from the above feature, the feature
in the text statement is "0." The shortest dependency path of the
entity pair in the text sentence is eight words between the entity
pairs. If the extracted feature is different from the above
feature, the feature in the text statement is "0."
In the embodiment, the above lexical features and syntactic
features are used as first features and second features. It is easy
to determine the features and simple to apply the method.
Another example of the present disclosure is an information
extracting apparatus. As shown in FIG. 7, the information
extracting apparatus comprises an acquisition module 10 and a
screening module 20.
The acquisition module 10 is configured to acquire entity pairs
corresponding to given semantic relationships and first instances
corresponding to the entity pairs. The first instance is a sentence
containing the entity pair.
The screening module 20 is configured to remove instances that do
not have the given semantic relationships from the first instances
to obtain second instances.
In an information extraction apparatus according to one embodiment,
the acquisition module 10 acquires entity pairs corresponding to
given semantic relationships. According to the entity pairs, the
first instances are obtained. Then, the screening module 20 screens
the first instances, and removes the first instances that do not
have the given semantic relationship. The first instances having
the given semantic relationship are screened out to form second
instances, thereby removing noise data from sample data. In this
way, when the sample data is used to train a model, accuracy of the
model is improved. Accordingly, accuracy of the acquired
information is improved.
In one embodiment, the acquisition module 10 comprises: acquiring
entity pairs corresponding to given semantic relationships from a
knowledge base, tagging sentences containing the named entities in
the knowledge base using a named entity recognition tool, and
retrieving first instances containing the entity pairs from the
tagged sentences. In the embodiment of the present disclosure, the
acquisition module 10 acquires the first instances containing the
entity pairs from the database by using the named entity
recognition tool. As such, efficiency is high, and cost is low.
In one embodiment, the screening module 20 comprises: extracting
first features from each of the first instances according to the
above-described entity pairs to construct a first instance-feature
matrix M.sub.nf, constructing a semantic relationship-first
instance matrix M.sub.rn, constructing a semantic
relationship-feature matrix M.sub.rf, wherein
M.sub.rf=M.sub.rn*M.sub.nf, decomposing the semantic
relationship-feature matrix M.sub.rf into M.sub.rk*M.sub.kf by a
nonnegative matrix factorization method, wherein k<f, obtaining
M.sub.nk by multiplying M.sub.nf by M.sub.kf.sup.T, obtaining
similarity between each of the first instances and the semantic
relationships respectively based on Mnk and Mrk.sup.T; and
screening out the second instances from the first instances based
on the similarity.
n is the number of the first instances. f is the number of all the
first features. The first features comprise lexical features and/or
syntactic features. r is the number of the semantic relationships.
M.sub.rk is a representation matrix of the semantic relationships
in a latent semantic space. M.sub.nk is a representation matrix of
the first instance in the latent semantic space;
In the embodiment of the present disclosure, the screening module
20 first maps the semantic relationships, instances, and features
into the latent semantic space by using a nonnegative matrix
factorization method. Then, similarity of each of the first
instances and the semantic relationships respectively in the latent
semantic space is obtained. Then, second instances are screened out
from the first instances based on whether the similarity has a
significant distinction, that is, whether it can be clearly
determined from the similarity that an instance belongs to a
semantic relationship. For this data retrieval method based on
matrix decomposition, the amount of calculation is small and
efficiency is high during the process of removing noise data.
In one embodiment, in the screening module 20, obtaining similarity
between each of the first instances and the semantic relationships
respectively based on M.sub.nk and M.sub.rk.sup.T comprises:
obtaining similarity between each of the first instances and the
semantic relationships respectively through cosine similarity based
M.sub.nk and M.sub.rk.sup.T. In the embodiment, the similarity
between each of the first instances and the semantic relationships
respectively is obtained by calculating the cosine similarity. This
embodiment obtains excellent result, has quick speed, and is
convenient and simple.
In one embodiment, in the screening module 20, screening out the
second instances from the first instances based on the similarity
comprises: normalizing the similarity so that a sum of the
similarities of each of the first instances to all semantic
relationships respectively is 1; calculating information entropy of
each of the first instances based on the corresponding normalized
similarities of each of the first instances; and selecting the
first instances whose information entropy is smaller than a
predetermined threshold as the second instances.
The embodiment of the present disclosure normalizes the similarity
so that the sum of the similarities of each of the first instance
to all semantic relationships respectively is 1. Then, the
similarity is regarded as a probability distribution to calculate
the information entropy of each of the instances. Then, noise data
is eliminated by selecting instances that satisfy the predetermined
threshold of the information entropy. As such, the second instances
are selected from the first instances. The calculation amount in
this embodiment is small and the acquisition thereof is
convenient.
In one embodiment, as shown in FIG. 7, the apparatus further
comprises a classifier training module 30 and an information
extraction module 40.
The classifier training module 30 is configured to extract first
features from each of second instances based on the entity pairs to
train the classifier. The target of the classifier is to classify
based on given semantic relationships. The first features comprise
lexical features and/or syntactic features.
The information extraction module 40 is configured to identify a
given semantic relationship from a text sentence and classify it
according to the classifier.
In the embodiment of the present disclosure, the classifier
training module 30 employs the second instances (sample data) in
which the noise data is removed to train the classifier. As such,
parameters of the classifier trained by the second instances are
more desirable. When the information extraction module 40 is used
to extract information of a text sentence, accuracy of the
extracted information is improved.
In addition, the present embodiment uses a trained classifier to
supervise and learn from an existing database from long-distance
and to identify and classify the semantic relationship based on the
entity pair in the text sentence automatically, instead of manually
tagging data. As such, cost of tagging is reduced, extraction
coverage of the semantic relationships is improved, and problem of
sparse data is solved.
In one embodiment, the information extraction module 40 comprises:
tagging sentences that contain named entities in a database by a
named entity recognition tool; retrieving text sentences containing
the entity pairs from the tagged sentences: and identifying the
corresponding semantic relationships of the entity pairs in the
text sentences and classifying the text sentences based on the
semantic relationships according to the classifier.
In the embodiment, the information extraction module 40 tags
sentences containing named entities in the database using the named
entity recognition tool. Then, the text sentences containing the
entity pairs are retrieved from the tagged sentences. Then, the
classifier is used to identify the corresponding semantic
relationships of the entity pairs in the text sentences and
classify the text sentences based on the semantic relationships. As
such, cost is low and efficiency is high.
In one embodiment, the lexical features include at least one of
entity pair's position in a sentence, word sequence between the
entity pair, characteristic sequence between the entity pair, left
window of a size of X of the entity pair, or right window of a size
of X of the entity pair, wherein X.gtoreq.0.
The syntax feature includes at least one of the shortest dependency
path between the entity pairs, left window of a size of Y of the
entity pair, or right window of a size of Y of the entity pair,
wherein Y.gtoreq.0.
The embodiment of the present disclosure uses the above lexical
features and syntactic features as first features and second
features. As such, it is easy to determine the features and simple
to apply the method.
The descriptions of the various embodiments of the present
disclosure have been presented for purposes of illustration, but
are not intended to be exhaustive and the limitation is not limited
to the embodiments disclosed. Many modifications and variations
will be apparent to those of ordinary skill in the art without
departing from the scope and spirit of the described embodiments.
The terminology used herein was chosen to best explain the
principles of the embodiments, the practical application or
technical improvement over technologies found in the marketplace,
or to enable others of ordinary skill in the art to understand the
embodiments disclosed herein.
REFERENCE OF THE FIGURES
10--acquisition module 20--filter module 30--classifier training
module 40--information extraction module
* * * * *